PICCA Interactive Explorer

An interactive artifact for Cross-Phase Blending: Why AI Fabrics Need Phase-Aware Transport. Explore how scalar congestion-control history differs from phase-indexed history under recurring AI communication phases with jitter, finite feedback delay, and increasing link speed.

Watch the longer PICCA video

Controls

Traffic model: four recurring communication phases. Their ideal rates differ, and the duration of each phase jitters from iteration to iteration.
Feedback and controller parameters. A scalar controller carries one running history across all phases. PICCA stores one history per phase and revisits the appropriate slot.
Boundary-cost lens. For a normalized state error Δx, stale-state bytes scale roughly as Δx × C × RTT. Here we visualize this with a nominal message size and line rate.
Preset scenarios

Simulation

Raw trace Ideal phase rate Scalar history PICCA / phase-indexed Stale boundary windows
PhaseInterpretationIdeal normalized rate
p1Large collective / heavy communication0.90
p2Low-demand phase / overlap gap0.10
p3Moderate communication phase0.50
p4Another high-demand phase0.80